[
https://issues.apache.org/jira/browse/HDFS-9075?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14746675#comment-14746675
]
He Tianyi commented on HDFS-9075:
---------------------------------
Thanks for point that out, [~cnauroth].
Prior discusses mentioned global namespace model, which i think is the most
valuable direction to work on.
There are consistency choices about namespace:
1. strong consistent namespace, perhaps either requires a global quorum to
ensure consistency or namespace segmentation (bit like federation, with only
local block pool)
2. eventual consistent namespace, can be achieved via snapshots. (What happens
Besides, there are choices about data replication fashion:
1. sync replication, add remote nodes to pipeline during write,
2. async replication.
IMHO strong consistent namespace is a must, otherwise global operations tend to
be hard to become transparent.
i.e. What happens if append operation on different file (or same file) in same
directory take place simultaneously in two datacenters?
(Of course a global lease manager would do the trick, but that requires remote
communication)
If we go the strong consistent way, performance suffers anyway (R/W needs
global communication anyway). It's no harm simply use one central active
NameNode, but with JournalNode and standby NameNode deployed globally.
As for replication, I think performance will not be an issue when given latency
is tolerable and bandwidth is sufficient (See HDFS-8829). We can certainly let
user decide.
We have a real scenario that communication between two datacenters have a
latency of nearly 3ms, while bandwidth is sufficient.
In this case, we see no performance drop so far.
But with high latency, I think that will not hold. Perhaps we need some fresh
idea.
> Multiple datacenter replication inside one HDFS cluster
> -------------------------------------------------------
>
> Key: HDFS-9075
> URL: https://issues.apache.org/jira/browse/HDFS-9075
> Project: Hadoop HDFS
> Issue Type: New Feature
> Components: datanode, namenode
> Reporter: He Tianyi
> Assignee: He Tianyi
>
> It is common scenario of deploying multiple datacenter for scaling and
> disaster tolerant.
> In this case we certainly want that data can be shared transparently (to
> user) across datacenters.
> For example, say we have a raw user action log stored daily, different
> computations may take place with the log as input. As scale grows, we may
> want to schedule various kind of computations in more than one datacenter.
> As far as i know, current solution is to deploy multiple independent clusters
> corresponding to datacenters, using {{distcp}} to sync data files between
> them.
> But in this case, user needs to know exactly where data is stored, and
> mistakes may be made during human-intervened operations. After all, it is
> basically a computer job.
> Based on these facts, it is obvious that a multiple datacenter replication
> solution may solve the scenario.
> I am working one prototype that works with 2 datacenters, the goal is to
> provide data replication between datacenters transparently and minimize the
> inter-dc bandwidth usage. Basic idea is replicate blocks to both DC and
> determine number of replications by historical statistics of access behaviors
> of that part of namespace.
> I will post a design document soon.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
